Fast Whole-Brain MR Multi-Parametric Mapping with Scan-Specific Self-Supervised Networks

Abstract

Quantification of tissue parameters using MRI is emerging as a powerful tool in clinical diagnosis and research studies. The need for multiple long scans with different acquisition parameters prohibits quantitative MRI from reaching widespread adoption in routine clinical and research exams. Accelerated parameter mapping techniques leverage parallel imaging, signal modelling and deep learning to offer more practical quantitative MRI acquisitions. However, the achievable acceleration and the quality of maps are often limited. Joint MAPLE is a recent state-of-the-art multi-parametric and scan-specific parameter mapping technique with promising performance at high acceleration rates. It synergistically combines parallel imaging, model-based and machine learning approaches for joint mapping of $T_1,{T_2}^{*}$ , proton density and the field inhomogeneity. However, Joint MAPLE suffers from prohibitively long reconstruction time to estimate the maps from a multi-echo, multi-flip angle (MEMFA) dataset at high resolution in a scan-specific manner. In this work, we propose a faster version of Joint MAPLE which retains the mapping performance of the original version. Coil compression, random slice selection, parameter-specific learning rates and transfer learning are synergistically combined in the proposed framework. It speeds-up the reconstruction time up to 700 times than the original version and processes a whole-brain MEMFA dataset in 21 minutes on average, which originally requires ~260 hours for Joint MAPLE. The mapping performance of the proposed framework is ~2-fold better than the standard and the state-of-the-art evaluated reconstruction techniques on average in terms of the root mean squared error.

Keywords: parameter mapping; quantitative MRI; scan-specific deep learning; self-supervised networks.

Fig. 1.. Joint MAPLE framework.

Joint ZS-SSL reconstructs under-sampled multi-echo, multi-flip angle k-space data (recon block) and multi-parameter signal model generates synthetic images using the unknown MR maps (model block). Matching reconstructed and synthesized contrasts form ${Loss}_1$ and matching k-space related to the synthetic images (using the forward model A) and acquired data forms ${Loss}_2$. The unknown MR parameters are estimated in a plug-and-play manner between recon and model block with minimization of Loss function, which is a scaled summation of ${Loss}_1,{Loss}_2$ and the training loss term of the joint ZS-SSL network ${Loss}_R$.

Fig. 2.. Joint ZS-SSL overall framework for whole-brain reconstruction.

Joint ZS-SSL is extended for whole-brain reconstruction of coil-compressed data with a random-slice selection manner. The joint ZS-SSL unrolls the ResNet-DC units with integrated convolution of different contrasts in the deep layers. It splits the acquired original k-space mask into training, loss and validation masks which are complementary across the contrasts. In each training iteration a random slice is selected to be fed into the network and the stopping criteria is extended to validate the generalizability of the network across different slices of a volume.

Fig. 3.. FTL-Joint MAPLE vs parallel imaging techniques.

The performance of the proposed FTL-Joint MAPLE vs SENSE, LORAKS and Joint MAPLE in terms of whole-brain reconstruction time (WBRT) and the showcased single slice normalized RMSE. The acceleration rate is R = 12 (4×3) with a complementary uniform sampling scheme. The maps and the metrics for SENSE and LORAKS are the average across TEs/FAs. Reported WBRT for FTL-Joint MAPLE is the average of 10 runs. The proposed FTL-Joint MAPLE is significantly faster than the original Joint MAPLE with close NRMSEs in all parameters. It also outperforms SENSE and LORAKS significantly in both accuracy and reconstruction time for all parameters.

Fig. 4.. Generalizability evaluation.

Generalizability comparison of Joint MAPLE trained by slice 15 and FTL-Joint MAPLE trained by the random slice selection manner. The acceleration rate is R = 12 (4×3) and a complementary uniform sampling is used. For slice 15, two methods perform closely. For slices 30 and 42 which are unseen for Joint MAPLE, the mapping performance of the proposed FTL-Joint MAPLE is better than Joint MAPLE in all parameters which is an indicator of better obtained generalizability with whole-brain training.

Fig. 5.. Mapping performance of FTL-Joint MAPLE vs SENSE, joint ZS-SSL and LORAKS for different slices.

The acceleration rate is R = 12(4×3) with a complementary uniform sampling scheme. The results for SENSE, joint ZS-SSL and LORAKS are the average results across TEs/FAs. FTL-Joint MAPLE outperforms other techniques for different slices in all parameters and it retains its mapping performance while accelerating the reconstruction process. The reconstruction of lower slices is more challenging for all techniques.

Fig. 6.. A three-view (axial, sagittal and coronal) of the whole-brain reconstructed T1 and T2* maps which showcases the middle slice of each view.

The acceleration rate is R = 12 (4×3) with a complementary uniform pattern. The reported NRMSEs are the whole-brain NRMSEs after applying a brain mask.

Fig. 7.. Ablation study of the effect of each proposed contribution on the mapping metrics in successive parts.

The acceleration rate is R = 12 (4×3) with a complementary uniform pattern. A) The original Joint MAPLE with no acceleration contribution which shows the best overall mapping performance in terms of the measured errors. B) Joint MAPLE performance on the coil-compressed version of the same dataset. The significant reduction in reconstruction time shows the impact of the size of dataset in the run time of the framework. C) Incorporating whole-brain reconstruction and coil compression into Joint MAPLE reduces the reconstruction time from a few days to ~ 7 hours. D) Optimization of whole-brain Joint MAPLE with parameter-specific learning rates using coil-compressed dataset which decreases the reconstruction time to less than one hour. E) The proposed FTL-Joint MAPLE with added transfer learning with the best reconstruction time and mapping performance comparable with the original Joint MAPLE (part A).

Fig. 8.. Synthetic images generated by parallel imaging techniques vs FTL-Joint MAPLE.

Eq. (3) is used for SENSE, joint ZS-SSL and LORAKS, and FTL-Joint MAPLE uses Eq. (2) as the signal model. The generated contrast related to FA = 10° and TE = 3.6ms for slice 30 is showcased with 10x scaled error image for each technique. NRMSEs are the average values across different TEs/FAs of slice 30. While its outperformance in MR parameter mapping, the synthetic images generated by FTL-Joint MAPLE show higher error values due to the model mismatch effect.